SPARK: Stable and Efficient Training for Spiking Neural Networks

Modern AI, particularly with large Foundation Models, achieves high capabilities at the cost of unsustainable energy consumption. The SURE-AI Center correctly identifies this dependance on overparameterization and vast datasets as an unsustainable practice. Spiking Neural Networks (SNNs) offer a biologically inspired alternative, mimicking the brain's sparse, temporal communication for unparalleled energy efficiency, providing a clear path toward the sustainable AI systems that the Center targets. However, SNN adoption is critically hindered by a training bottleneck: the non-differentiable spiking activation prevents standard backpropagation. Current heuristics, such as the straight-through estimator (STE), are mathematically tenuous, leading to unstable training and a significant performance gap. This reliance causes a critical lack of rigor and performance guarantees. Furthermore, existing training methods fail to fully exploit the asynchronous, temporal dynamics, which is the true source of SNN efficiency, and this limits their scalability. This ongoing bottleneck prevents SNNs from reaching their full potential. This project proposes the design of a next-generation framework to provide a mathematically rigorous, stable, and highly effective SNN training methodology.

By focusing on a theoretically sound gradient approximation, we will overcome heuristic performance limitations and will provide performance guarantees. The project will bridge SNN theory with practical deployment, validating the framework on cutting-edge neuromorphic hardware to provide empirical proof of superior energy efficiency, enabling the widespread adoption of truly sustainable and trustworthy AI systems within the SURE-AI center and beyond.

Funding

Funded by the Research Council of Norway through the "Coordination and Support Activity Support for Researcher Mobility" scheme.